Professional Documents
Culture Documents
net/publication/237078258
CITATIONS READS
4 49
2 authors:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Piergiorgio Lovaglio on 24 January 2014.
1
This paper is the result of the joint effort of both the authors. G. Vittadini took care of
the final editing of Section 1, whereas P.G. Lovaglio was responsible for all the other
Sections.
356 P. G. Lovaglio, G. Vittadini
Internal External
In the gain score approach the “adjusted outcome” is simply the expected
value of gain, E(y2-y1), whereas the added value or residual gain score ap-
proach2 is concerned with the quantity y2-E(y2/y1), i.e. the difference between
the post-test and its expected value E(y2/y1), a result of the regression of y2
onto y1.
Both approaches present some drawbacks. Some researchers highlight the
technical problems associated with the residual gain score approach (Cronbach
& Furby, 1970; Scheerens & Bosker, 1997; Hanushek, 1997, 2002; Willett,
1988; Thum, 2002). First, it is reasonable to expect that the same set of factors
affecting the pre-test would affect the post-test as well. Like the post-test, the
pre-test is measured with error, so that a regression of post-test on a pre-test
score violates the basic regression assumption that covariates are error-free.
This implies that the residual gain score is an inefficient and inconsistent esti-
mator of true change. Secondly, because the pre-test typically explains a large
quota of the post-test outcome, this drastic reduction of residual variance lim-
its the explicative power of other factors, for example the resources of the or-
ganization, its quality and so on.
Other researchers (Rogosa et al., 1982; Rogosa & Willet, 1985; Rogosa,
1995; Singer & Willet, 2003) dispel the notion that gain scores are inherently
unreliable, so demonstrating that the correlation between gain and initial status
(the major drawback of the gain score approach) is merely an artefact of the
design. The nature of the correlation is the outcome of the parameterization we
choose for the growth factors and, as a consequence, the meaning itself we at-
tach to them.
Thum (2002) and Bryk et al. (1998) propose to measure the outcomes in
time series in a longitudinal framework, an approach that allows insertion of
covariates related to the institutions that can affect the performances of the
micro units (students, classes). The repeated measurement approach we adopt
in this paper places the pre-test and the post-test on an equal footing in that
both serve as outcomes in our models. The change with respect to gain scores
is evident in recent research (Collins, 1996; Maris, 1998; Mellenbergh, 1999;
Williams & Zimmerman, 1996).
The second development, known as “hierarchical", “multilevel", “growth”
or “mixed-effects" modelling, provides the methodological complement for a
proper, flexible, treatment of growth data in stratified sampling designs that
are common in educational research settings (repeated measures on students
over time, students nested with schools in the district, etc.).
2
Its origins can be traced to notions of goods, value and wealth in political economy
(Saunders, 1999); value-added analyses address a model for comparison (Goldstein &
Spiegelhalter, 1996) adjusted for the factors out of control for the researchers. Research-
ers have frequently applied the regression analysis of the post-test score on the pre-test
score, motivating this choice with the notion of “fairness" in order to flatten the playing
ground (Good et al. 1975).
358 P. G. Lovaglio, G. Vittadini
where: ytij represents the educational outcome at time t (t=1, .., k) for the i-th
subject (i = 1,…, nj) nested in the j–th institution (university, school, type of
programme) (j = 1,…, m); timetij is the variable time having cardinality equal
to the number of instants; π0ij and π1ij are the intercept (mean of outcome at
timetij=0) and slopes (rate of increase over time) for subject ij. In equation (1),
where for the sake of simplicity we did not insert covariates varying over time,
it is assumed that the outcome follows a linear trend.
We also specify a Level 2 (or between-person) model for inter-individual
differences. The Level 2 model decomposes the individual intercept and
slopes (π0ij and π1ij) into the mean values of the parameter institutions across
individuals (β00j and β10j), in the effect of a set of individual characteristics (xsij
with αsj associated parameter) of subject ij and the individual effect (u0ij and
u1ij).
Finally, the Level 3 model decomposes the mean effect of institutions into
the mean values of the parameters across all institutions (γ000 and γ100), in the
Human Capital Growth for University Education Evaluation 359
π 0ij = β00j + Σs αsjx sij + u 0ij π lij = β10j + Σsαsjx sij + u lij , (2)
Each equation presents errors at all nested levels: rtij is the random effect of
time t on the trajectory of outcome evolution for subject ij; u0ij and u1ij are the
random effects of subject ij with respect to the mean trajectory for the subjects
nested in institution j; m00j and m10j are the random effects of institution j with
respect to the mean trajectory for all institutions.
The distribution of the time specific disturbances, individual trajectory pa-
rameters and institutions’ specific parameters are assumed normally distrib-
uted, with zero expectation, and mutually independent:
rtij 0 Σ 0 0
(3)
u.ij ~ N 0 , 0 T 0
m.. j 0 0 0 G
where Σ is the (k,k) covariance matrix of the time-specific disturbances (rtij), T
the (2,2) covariance matrix of the individual trajectory parameters (u.ij) and G
the (2,2) covariance matrix of the trajectory parameters between institutions
residuals (m..j). Even if MGMs neglect the specification of the covariance be-
tween observations at different times on the same subject, a special attention
must be paid to Σ otherwise we risk incorrect conclusions on fixed effects
(Rogosa, 1995; Laird & Ware, 1982; Cnaan et al., 1997).
The model specified with formulas (1), (2) and (3) is applicable for estimat-
ing the specific effect of institutions on the level (in a temporal instant, m00j)
and on the rate of outcome increase over time (m10j). The model can be ex-
tended (Lovaglio, 2004) to the case of the evaluation of the efficiency among
institutions (relative evaluation) and to that of the distributed service (impact
evaluation).
In the latter case, the model allows to check if an institution that distributes
a service (higher education) obtains better performances in comparison with a
(control) group of subjects with similar characteristics to whom the service
was not distributed. To this aim, a dummy variable Iij – which assumes value 0
for subject i of the control group and 1 otherwise – is inserted in equations (2)
and (3) for the estimation of the impact level (κ0) and slope (κ1):
The proposed methodology was applied to the evaluation of the external effi-
ciency of Italian university education using the data collected by the Bank of
Italy with the Survey on Household Income and Wealth (Banca d’Italia, 2002)
for the years 1998, 2000 and 2002.
The dependent variable is the earned income for each individual in a time
series and the covariates are individual factors (social-demographic-economic
background), university covariates (input, resources, intangible aspects, per-
sonnel) and job market variables (region, province, etc. where the subject is
addressed).
Several indicators concur to identify the individual human capital and
household socio-economical status: Age, Gender, Region of residence, Marital
status, Educational level, Employment status, Type of job, Economic activity
sector, Years of schooling, Age of entrance in the labour market, Number of
children; Years of full-time work, Years part-time work, Household total
wealth, Household total debt, Household income, and Parents’ educational
level.
The outcome of the analysis is the partial disposable earned income (from
here on earned income) composed of earnings of employed and self-employed
people, pensions, transfers and economic assistance.
The sample was composed of subjects belonging to the labour force with
positive earned income for all three years; we identified the graduates as indi-
viduals who had achieved a university bachelor or master degree before 1998.
The panel of income earners is composed of less than 2000 subjects,
whereas the group of graduates of 559 individuals (Table 2). Because of the
reduced number of panel graduates, it is not possible to deepen our analyses
on the external efficiency by university, or type of degree. The analysis is
therefore limited to the effect of the higher education tout court on the income
dynamics (impact analysis that compares the level and the growth rates of
graduates and non-graduates earnings in Italy)3.
3
Such data moreover prevent the estimation of the short-term university education effect,
because we must select the graduates who have recently acquired a degree, typically go-
ing back three to five years from the first survey.
Human Capital Growth for University Education Evaluation 361
Table 3. Means of the earned income (in Euro) by degree level and time
Akaike's information criterion and Schwarz's Bayesian criterion are the in-
dices of relative goodness-of-fit chosen for comparing models with the same
fixed effects but different covariance structures. The covariance structure as-
sumed for Σ is the first order autoregressive one (the estimated correlation pa-
rameter for unitary lag is 0.373), which specifies homogeneous variance and
covariances between observations that decrease toward zero with increasing
lag. The test of heterogeneity indicates a common covariance structure for
graduates and non-graduates.
Once Σ is structured, the longitudinal analysis requires the construction of
the Unconditional Means Model (UMM), a model that does not require co-
variates at each level (level 1, level 2). It is useful to both estimate whether a
systematic variability exists in the outcome, and show how the total variance
of the outcome is decomposed between and within subjects, depending on the
time or on individual factors that changed over time.
The estimated intraclass correlation (the ratio of the between and total vari-
ances) shows that the 43% of the total variability of income is attributable to
the differences between the individuals (i.e. individual characteristics) and for
the remaining part to the differences within the subjects, confirming the hier-
archical structure (multilevel) of the dataset.
Inserting the variable time (as a continuous quantitative covariate) in the
UMM we obtain the Unconditional Growth Model (UGM) that explains how
much of the within subject variance is explained by time. We estimated that
28% of the within-subject variability is determined by time and the remaining
part by individual characteristics that change over time (occupation, sector, etc).
Relative to the fixed effects, the overall growth rate of earned income is
positive and highly significant, so indicating a strong dynamic effect on the
income trajectories, while, relatively to random parameters, the variance of in-
Human Capital Growth for University Education Evaluation 363
come in 1998 (the reference time) and the variance of the growth rates -
between the subjects- are significantly different from zero. The estimated co-
variance between the levels of incomes in 1998 and growth rates is not signifi-
cant, and this implies unstable relations between slopes and initial income.
In order to explain the residual variability between the levels (earned in-
comes in 1998) and between the growth rates (yearly increment of income),
we inserted individual covariates in the 2 level model (equation 5): some of
which, for the sake of simplicity, are kept fixed in time and measured at the fi-
nal year (2002).
The covariates and the significance levels of the associated parameters
(fixed effects)4 are shown in Table 5. The level of earned income at 1998 (left
part of Table 5), gender, type of occupation at 2002 (Occupation), possession
of a bachelor’s degree (Bachelor), years of full-time job (Full_job), occupa-
tional sector at 2002 (Sector), area of residence at 2002 (Area), marital status
at 2002 (Marital), and age at 2002 (Cl_age) are highly significant.
The factors that characterise the evolution of the growth rates are (right part
of Table 5): type of occupation at 2002 (t*Occupation), age at 2002
(t*Cl_age), occupational sector at 2002 (t*Sector) and, to a lesser extent, the
possession of a bachelor’s degree (t*Bachelor), whereas the other covariates
are not significant at 5% level. Relatively to the random parameters (not
shown here), the residual variances in the 1998 level and growth rates remain
significant.
We can conclude that income level at 1998 and income growth rates are
greater and faster increasing (even though close to the limits of the customary
significance levels) for graduates. The selected covariates explain a large
quota (56%) of the UGM residual variance for the level and just a part (21%)
of the between-subject growth rate variability, thus leaving open the search of
other individual covariates that may affect the evolution of income over time.
The trajectories estimated for each group of subjects (after adjustment for the
selected covariates) give a mean growth annual rates of 1,051€ for a graduate
and 426€ for a non-graduate.
Moreover, we estimated the means of growth rates for the two groups for
the more relevant age class. By comparing the growth rates of a graduate
(1507€) and a non-graduate (916€) in the age class less of 30 years it is possi-
ble to quantify how the labour market rewards university education. The im-
pact of a university degree appears highly significant in absolute terms, while
in comparison with the non-graduate rates this difference remains significant,
but constant in the older age groups (606€ for 31-40 year group; 584€ for 41-
50 group).
Finally, we estimated the differences between group income means at every
time and for the income growth rates. The contrast analysis shows that for all
4
The application of QQ plot test for Level 1 and Level 2 residuals shows that normality is
supported by data.
364 P. G. Lovaglio, G. Vittadini
22000
20000
18000
16000
14000
12000
1998 2000 2002
three temporal moments the differences between income means are not only
meaningful, but increasing in time (in particular, the difference is 1188€,
2015€ and 2118€, respectively for 1998, 2000 and 2002) and that the income
growth rate of the graduates (estimate: 525€) is larger than that of non-
graduates (p-value=0.0169).
Figure 1 shows the incomes trajectories for graduates and non-graduates,
relatively to the average observed means (Oss) and to the average estimated
means (Adj). The continuous lines, which show the income trends adjusted for
the subjects’ characteristics, confirm our previous results. The adjusted trajec-
tories make it evident the "correction" effect (also called risk-adjustment).
4. Conclusions
References
ROGOSA D.R. (1995) Myths about longitudinal research. In: GOTTMAN J.M. (ed),
The Analysis of Change, LEA, Mahwah, NJ: 3-66.
ROGOSA D.R., BRAND D., ZIMOWSKI M. (1982) A growth curve approach to the
measurement of change, Psychological Bulletin, 90: 726-748.
ROGOSA D.R., WILLETT J.B. (1985) Understanding correlates of change by model-
ing individual differences in growth, Psychometrika, 50: 203-228.
SAUNDERS L. (1999) A brief history of educational "value added": how did we get
to where we are?, School Effectiveness and School Improvement, 10(2): 233-256.
SCHEERENS J., BOSKER R.J. (1997) The Foundations of Educational Effectiveness,
Pergamon, Oxford.
SINGER J.D., WILLET J.B. (2003) Applied Longitudinal Data Analysis: Modeling
Change and Event Occurrence, Oxford University Press, Oxford.
THUM Y.M. (2002) Measuring progress towards a goal: estimating teacher productiv-
ity using a multivariate multilevel model for value-added analysis
(http://www.mff.org/pubs/tchprodv3.pdf).
WILLETT J.B. (1988) Questions and answers in the measurement of change, Review
of Research in Education, 15: 345-422.
WILLIAMS R.H., ZIMMERMAN D.W. (1996) Are gain scores obsolete?, Applied
Psychological Measurement, 20: 59-69.